Local-aware coupled network for hyperspectral image super-resolution

Abstract

Despite the unprecedented success of super-resolution (SR) development for natural images, achieving hyperspectral image (HSI) SR with rich spectral characteristics remains a challenging task. Typically, HSI SR is accomplished by fusing low-resolution HSI (LR HSI) with the corresponding high-resolution multispectral image (HR MSI). However, due to the significant spectral difference between MSI and HSI, it is difficult to retain the spatial characteristics of MSI during image fusion. In addition, the spectral response function (SRF) used for simulating MSI is often unknown or unavailable in hyperspectral remote sensing images, further complicating the problem. To address the above issues, a local-aware coupled network (LCNet) is proposed in this paper. In LCNet, the SRF and point spread function (PSF) are adaptively learned in the primary stage of the network to address the issue of unknown prior information. By coupling two reconstruction networks, LCNet effectively preserves both the texture details of MSI and the spectral characteristics of HSI. Furthermore, the spatial local-aware block selectively emphasizes the texture features of MSI. Experimental results on three publicly available HSIs demonstrate whether the proposed LCNet is superior to the state-of-the-art methods with respect to both stability and quality.